Parametric study of plasma swirling flow based on magnetohydrodynamics informed deep learning

A magnetohydrodynamics (MHD) informed deep learning framework is developed to address the challenges of achieving higher accuracy and faster optimization speeds in physics-informed neural networks (PINNs) for solving multi-parameter MHD problems without labeled data. By integrating hybrid boundary conditions, optimized activation functions, and residual-based adaptive sampling strategy, the proposed framework achieves a global relative error of approximately 2% across a wide parameter range. To address the challenge of training coupled equations in broad parameter regimes, we introduce coefficient normalization for coupled terms that balances loss contributions and prevents equation decoupling. Through this MHD-informed deep learning approach, comprehensive parametric studies and parameter optimizations are conducted to systematically investigate the effects of magnetic field strength, mass flow rate, and geometric configuration on plasma swirling flow dynamics. The proposed methodology leverages the automatic differentiability of PINNs for rapid parameter optimization, achieving up to an order-of-magnitude acceleration in parameter optimization compared to conventional methods like Genetic Algorithm while maintaining comparable accuracy. This advancement demonstrates significant potential for the parametric analysis and optimization design of MHD flow-based devices in the areas of advanced space thruster and manufacturing technologies.